fs/gfs2/aops.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License version 2.
   8  */
   9
  10 #include <linux/sched.h>
  11 #include <linux/slab.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/completion.h>
  14 #include <linux/buffer_head.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/pagevec.h>
  17 #include <linux/mpage.h>
  18 #include <linux/fs.h>
  19 #include <linux/writeback.h>
  20 #include <linux/swap.h>
  21 #include <linux/gfs2_ondisk.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/uio.h>
  24 #include <trace/events/writeback.h>
  25
  26 #include "gfs2.h"
  27 #include "incore.h"
  28 #include "bmap.h"
  29 #include "glock.h"
  30 #include "inode.h"
  31 #include "log.h"
  32 #include "meta_io.h"
  33 #include "quota.h"
  34 #include "trans.h"
  35 #include "rgrp.h"
  36 #include "super.h"
  37 #include "util.h"
  38 #include "glops.h"
  39
  40
  41 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  42                                    unsigned int from, unsigned int to)
  43 {
  44         struct buffer_head *head = page_buffers(page);
  45         unsigned int bsize = head->b_size;
  46         struct buffer_head *bh;
  47         unsigned int start, end;
  48
  49         for (bh = head, start = 0; bh != head || !start;
  50              bh = bh->b_this_page, start = end) {
  51                 end = start + bsize;
  52                 if (end <= from || start >= to)
  53                         continue;
  54                 if (gfs2_is_jdata(ip))
  55                         set_buffer_uptodate(bh);
  56                 gfs2_trans_add_data(ip->i_gl, bh);
  57         }
  58 }
  59
  60 /**
  61  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  62  * @inode: The inode
  63  * @lblock: The block number to look up
  64  * @bh_result: The buffer head to return the result in
  65  * @create: Non-zero if we may add block to the file
  66  *
  67  * Returns: errno
  68  */
  69
  70 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  71                                   struct buffer_head *bh_result, int create)
  72 {
  73         int error;
  74
  75         error = gfs2_block_map(inode, lblock, bh_result, 0);
  76         if (error)
  77                 return error;
  78         if (!buffer_mapped(bh_result))
  79                 return -EIO;
  80         return 0;
  81 }
  82
  83 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  84                                  struct buffer_head *bh_result, int create)
  85 {
  86         return gfs2_block_map(inode, lblock, bh_result, 0);
  87 }
  88
  89 /**
  90  * gfs2_writepage_common - Common bits of writepage
  91  * @page: The page to be written
  92  * @wbc: The writeback control
  93  *
  94  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
  95  */
  96
  97 static int gfs2_writepage_common(struct page *page,
  98                                  struct writeback_control *wbc)
  99 {
 100         struct inode *inode = page->mapping->host;
 101         struct gfs2_inode *ip = GFS2_I(inode);
 102         struct gfs2_sbd *sdp = GFS2_SB(inode);
 103         loff_t i_size = i_size_read(inode);
 104         pgoff_t end_index = i_size >> PAGE_SHIFT;
 105         unsigned offset;
 106
 107         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 108                 goto out;
 109         if (current->journal_info)
 110                 goto redirty;
 111         /* Is the page fully outside i_size? (truncate in progress) */
 112         offset = i_size & (PAGE_SIZE-1);
 113         if (page->index > end_index || (page->index == end_index && !offset)) {
 114                 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
 115                 goto out;
 116         }
 117         return 1;
 118 redirty:
 119         redirty_page_for_writepage(wbc, page);
 120 out:
 121         unlock_page(page);
 122         return 0;
 123 }
 124
 125 /**
 126  * gfs2_writepage - Write page for writeback mappings
 127  * @page: The page
 128  * @wbc: The writeback control
 129  *
 130  */
 131
 132 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 133 {
 134         int ret;
 135
 136         ret = gfs2_writepage_common(page, wbc);
 137         if (ret <= 0)
 138                 return ret;
 139
 140         return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
 141 }
 142
 143 /* This is the same as calling block_write_full_page, but it also
 144  * writes pages outside of i_size
 145  */
 146 static int gfs2_write_full_page(struct page *page, get_block_t *get_block,
 147                                 struct writeback_control *wbc)
 148 {
 149         struct inode * const inode = page->mapping->host;
 150         loff_t i_size = i_size_read(inode);
 151         const pgoff_t end_index = i_size >> PAGE_SHIFT;
 152         unsigned offset;
 153
 154         /*
 155          * The page straddles i_size.  It must be zeroed out on each and every
 156          * writepage invocation because it may be mmapped.  "A file is mapped
 157          * in multiples of the page size.  For a file that is not a multiple of
 158          * the  page size, the remaining memory is zeroed when mapped, and
 159          * writes to that region are not written out to the file."
 160          */
 161         offset = i_size & (PAGE_SIZE-1);
 162         if (page->index == end_index && offset)
 163                 zero_user_segment(page, offset, PAGE_SIZE);
 164
 165         return __block_write_full_page(inode, page, get_block, wbc,
 166                                        end_buffer_async_write);
 167 }
 168
 169 /**
 170  * __gfs2_jdata_writepage - The core of jdata writepage
 171  * @page: The page to write
 172  * @wbc: The writeback control
 173  *
 174  * This is shared between writepage and writepages and implements the
 175  * core of the writepage operation. If a transaction is required then
 176  * PageChecked will have been set and the transaction will have
 177  * already been started before this is called.
 178  */
 179
 180 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 181 {
 182         struct inode *inode = page->mapping->host;
 183         struct gfs2_inode *ip = GFS2_I(inode);
 184         struct gfs2_sbd *sdp = GFS2_SB(inode);
 185
 186         if (PageChecked(page)) {
 187                 ClearPageChecked(page);
 188                 if (!page_has_buffers(page)) {
 189                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 190                                              BIT(BH_Dirty)|BIT(BH_Uptodate));
 191                 }
 192                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 193         }
 194         return gfs2_write_full_page(page, gfs2_get_block_noalloc, wbc);
 195 }
 196
 197 /**
 198  * gfs2_jdata_writepage - Write complete page
 199  * @page: Page to write
 200  * @wbc: The writeback control
 201  *
 202  * Returns: errno
 203  *
 204  */
 205
 206 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
 207 {
 208         struct inode *inode = page->mapping->host;
 209         struct gfs2_inode *ip = GFS2_I(inode);
 210         struct gfs2_sbd *sdp = GFS2_SB(inode);
 211         int ret;
 212
 213         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
 214                 goto out;
 215         if (PageChecked(page) || current->journal_info)
 216                 goto out_ignore;
 217         ret = __gfs2_jdata_writepage(page, wbc);
 218         return ret;
 219
 220 out_ignore:
 221         redirty_page_for_writepage(wbc, page);
 222 out:
 223         unlock_page(page);
 224         return 0;
 225 }
 226
 227 /**
 228  * gfs2_writepages - Write a bunch of dirty pages back to disk
 229  * @mapping: The mapping to write
 230  * @wbc: Write-back control
 231  *
 232  * Used for both ordered and writeback modes.
 233  */
 234 static int gfs2_writepages(struct address_space *mapping,
 235                            struct writeback_control *wbc)
 236 {
 237         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
 238         int ret = mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 239
 240         /*
 241          * Even if we didn't write any pages here, we might still be holding
 242          * dirty pages in the ail. We forcibly flush the ail because we don't
 243          * want balance_dirty_pages() to loop indefinitely trying to write out
 244          * pages held in the ail that it can't find.
 245          */
 246         if (ret == 0)
 247                 set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
 248
 249         return ret;
 250 }
 251
 252 /**
 253  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
 254  * @mapping: The mapping
 255  * @wbc: The writeback control
 256  * @pvec: The vector of pages
 257  * @nr_pages: The number of pages to write
 258  * @end: End position
 259  * @done_index: Page index
 260  *
 261  * Returns: non-zero if loop should terminate, zero otherwise
 262  */
 263
 264 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
 265                                     struct writeback_control *wbc,
 266                                     struct pagevec *pvec,
 267                                     int nr_pages, pgoff_t end,
 268                                     pgoff_t *done_index)
 269 {
 270         struct inode *inode = mapping->host;
 271         struct gfs2_sbd *sdp = GFS2_SB(inode);
 272         unsigned nrblocks = nr_pages * (PAGE_SIZE/inode->i_sb->s_blocksize);
 273         int i;
 274         int ret;
 275
 276         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
 277         if (ret < 0)
 278                 return ret;
 279
 280         for(i = 0; i < nr_pages; i++) {
 281                 struct page *page = pvec->pages[i];
 282
 283                 /*
 284                  * At this point, the page may be truncated or
 285                  * invalidated (changing page->mapping to NULL), or
 286                  * even swizzled back from swapper_space to tmpfs file
 287                  * mapping. However, page->index will not change
 288                  * because we have a reference on the page.
 289                  */
 290                 if (page->index > end) {
 291                         /*
 292                          * can't be range_cyclic (1st pass) because
 293                          * end == -1 in that case.
 294                          */
 295                         ret = 1;
 296                         break;
 297                 }
 298
 299                 *done_index = page->index;
 300
 301                 lock_page(page);
 302
 303                 if (unlikely(page->mapping != mapping)) {
 304 continue_unlock:
 305                         unlock_page(page);
 306                         continue;
 307                 }
 308
 309                 if (!PageDirty(page)) {
 310                         /* someone wrote it for us */
 311                         goto continue_unlock;
 312                 }
 313
 314                 if (PageWriteback(page)) {
 315                         if (wbc->sync_mode != WB_SYNC_NONE)
 316                                 wait_on_page_writeback(page);
 317                         else
 318                                 goto continue_unlock;
 319                 }
 320
 321                 BUG_ON(PageWriteback(page));
 322                 if (!clear_page_dirty_for_io(page))
 323                         goto continue_unlock;
 324
 325                 trace_wbc_writepage(wbc, inode_to_bdi(inode));
 326
 327                 ret = __gfs2_jdata_writepage(page, wbc);
 328                 if (unlikely(ret)) {
 329                         if (ret == AOP_WRITEPAGE_ACTIVATE) {
 330                                 unlock_page(page);
 331                                 ret = 0;
 332                         } else {
 333
 334                                 /*
 335                                  * done_index is set past this page,
 336                                  * so media errors will not choke
 337                                  * background writeout for the entire
 338                                  * file. This has consequences for
 339                                  * range_cyclic semantics (ie. it may
 340                                  * not be suitable for data integrity
 341                                  * writeout).
 342                                  */
 343                                 *done_index = page->index + 1;
 344                                 ret = 1;
 345                                 break;
 346                         }
 347                 }
 348
 349                 /*
 350                  * We stop writing back only if we are not doing
 351                  * integrity sync. In case of integrity sync we have to
 352                  * keep going until we have written all the pages
 353                  * we tagged for writeback prior to entering this loop.
 354                  */
 355                 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
 356                         ret = 1;
 357                         break;
 358                 }
 359
 360         }
 361         gfs2_trans_end(sdp);
 362         return ret;
 363 }
 364
 365 /**
 366  * gfs2_write_cache_jdata - Like write_cache_pages but different
 367  * @mapping: The mapping to write
 368  * @wbc: The writeback control
 369  *
 370  * The reason that we use our own function here is that we need to
 371  * start transactions before we grab page locks. This allows us
 372  * to get the ordering right.
 373  */
 374
 375 static int gfs2_write_cache_jdata(struct address_space *mapping,
 376                                   struct writeback_control *wbc)
 377 {
 378         int ret = 0;
 379         int done = 0;
 380         struct pagevec pvec;
 381         int nr_pages;
 382         pgoff_t uninitialized_var(writeback_index);
 383         pgoff_t index;
 384         pgoff_t end;
 385         pgoff_t done_index;
 386         int cycled;
 387         int range_whole = 0;
 388         int tag;
 389
 390         pagevec_init(&pvec, 0);
 391         if (wbc->range_cyclic) {
 392                 writeback_index = mapping->writeback_index; /* prev offset */
 393                 index = writeback_index;
 394                 if (index == 0)
 395                         cycled = 1;
 396                 else
 397                         cycled = 0;
 398                 end = -1;
 399         } else {
 400                 index = wbc->range_start >> PAGE_SHIFT;
 401                 end = wbc->range_end >> PAGE_SHIFT;
 402                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
 403                         range_whole = 1;
 404                 cycled = 1; /* ignore range_cyclic tests */
 405         }
 406         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 407                 tag = PAGECACHE_TAG_TOWRITE;
 408         else
 409                 tag = PAGECACHE_TAG_DIRTY;
 410
 411 retry:
 412         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
 413                 tag_pages_for_writeback(mapping, index, end);
 414         done_index = index;
 415         while (!done && (index <= end)) {
 416                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
 417                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
 418                 if (nr_pages == 0)
 419                         break;
 420
 421                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end, &done_index);
 422                 if (ret)
 423                         done = 1;
 424                 if (ret > 0)
 425                         ret = 0;
 426                 pagevec_release(&pvec);
 427                 cond_resched();
 428         }
 429
 430         if (!cycled && !done) {
 431                 /*
 432                  * range_cyclic:
 433                  * We hit the last page and there is more work to be done: wrap
 434                  * back to the start of the file
 435                  */
 436                 cycled = 1;
 437                 index = 0;
 438                 end = writeback_index - 1;
 439                 goto retry;
 440         }
 441
 442         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
 443                 mapping->writeback_index = done_index;
 444
 445         return ret;
 446 }
 447
 448
 449 /**
 450  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
 451  * @mapping: The mapping to write
 452  * @wbc: The writeback control
 453  *
 454  */
 455
 456 static int gfs2_jdata_writepages(struct address_space *mapping,
 457                                  struct writeback_control *wbc)
 458 {
 459         struct gfs2_inode *ip = GFS2_I(mapping->host);
 460         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 461         int ret;
 462
 463         ret = gfs2_write_cache_jdata(mapping, wbc);
 464         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
 465                 gfs2_log_flush(sdp, ip->i_gl, NORMAL_FLUSH);
 466                 ret = gfs2_write_cache_jdata(mapping, wbc);
 467         }
 468         return ret;
 469 }
 470
 471 /**
 472  * stuffed_readpage - Fill in a Linux page with stuffed file data
 473  * @ip: the inode
 474  * @page: the page
 475  *
 476  * Returns: errno
 477  */
 478
 479 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 480 {
 481         struct buffer_head *dibh;
 482         u64 dsize = i_size_read(&ip->i_inode);
 483         void *kaddr;
 484         int error;
 485
 486         /*
 487          * Due to the order of unstuffing files and ->fault(), we can be
 488          * asked for a zero page in the case of a stuffed file being extended,
 489          * so we need to supply one here. It doesn't happen often.
 490          */
 491         if (unlikely(page->index)) {
 492                 zero_user(page, 0, PAGE_SIZE);
 493                 SetPageUptodate(page);
 494                 return 0;
 495         }
 496
 497         error = gfs2_meta_inode_buffer(ip, &dibh);
 498         if (error)
 499                 return error;
 500
 501         kaddr = kmap_atomic(page);
 502         if (dsize > (dibh->b_size - sizeof(struct gfs2_dinode)))
 503                 dsize = (dibh->b_size - sizeof(struct gfs2_dinode));
 504         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
 505         memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
 506         kunmap_atomic(kaddr);
 507         flush_dcache_page(page);
 508         brelse(dibh);
 509         SetPageUptodate(page);
 510
 511         return 0;
 512 }
 513
 514
 515 /**
 516  * __gfs2_readpage - readpage
 517  * @file: The file to read a page for
 518  * @page: The page to read
 519  *
 520  * This is the core of gfs2's readpage. Its used by the internal file
 521  * reading code as in that case we already hold the glock. Also its
 522  * called by gfs2_readpage() once the required lock has been granted.
 523  *
 524  */
 525
 526 static int __gfs2_readpage(void *file, struct page *page)
 527 {
 528         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 529         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 530         int error;
 531
 532         if (gfs2_is_stuffed(ip)) {
 533                 error = stuffed_readpage(ip, page);
 534                 unlock_page(page);
 535         } else {
 536                 error = mpage_readpage(page, gfs2_block_map);
 537         }
 538
 539         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 540                 return -EIO;
 541
 542         return error;
 543 }
 544
 545 /**
 546  * gfs2_readpage - read a page of a file
 547  * @file: The file to read
 548  * @page: The page of the file
 549  *
 550  * This deals with the locking required. We have to unlock and
 551  * relock the page in order to get the locking in the right
 552  * order.
 553  */
 554
 555 static int gfs2_readpage(struct file *file, struct page *page)
 556 {
 557         struct address_space *mapping = page->mapping;
 558         struct gfs2_inode *ip = GFS2_I(mapping->host);
 559         struct gfs2_holder gh;
 560         int error;
 561
 562         unlock_page(page);
 563         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 564         error = gfs2_glock_nq(&gh);
 565         if (unlikely(error))
 566                 goto out;
 567         error = AOP_TRUNCATED_PAGE;
 568         lock_page(page);
 569         if (page->mapping == mapping && !PageUptodate(page))
 570                 error = __gfs2_readpage(file, page);
 571         else
 572                 unlock_page(page);
 573         gfs2_glock_dq(&gh);
 574 out:
 575         gfs2_holder_uninit(&gh);
 576         if (error && error != AOP_TRUNCATED_PAGE)
 577                 lock_page(page);
 578         return error;
 579 }
 580
 581 /**
 582  * gfs2_internal_read - read an internal file
 583  * @ip: The gfs2 inode
 584  * @buf: The buffer to fill
 585  * @pos: The file position
 586  * @size: The amount to read
 587  *
 588  */
 589
 590 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
 591                        unsigned size)
 592 {
 593         struct address_space *mapping = ip->i_inode.i_mapping;
 594         unsigned long index = *pos / PAGE_SIZE;
 595         unsigned offset = *pos & (PAGE_SIZE - 1);
 596         unsigned copied = 0;
 597         unsigned amt;
 598         struct page *page;
 599         void *p;
 600
 601         do {
 602                 amt = size - copied;
 603                 if (offset + size > PAGE_SIZE)
 604                         amt = PAGE_SIZE - offset;
 605                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
 606                 if (IS_ERR(page))
 607                         return PTR_ERR(page);
 608                 p = kmap_atomic(page);
 609                 memcpy(buf + copied, p + offset, amt);
 610                 kunmap_atomic(p);
 611                 put_page(page);
 612                 copied += amt;
 613                 index++;
 614                 offset = 0;
 615         } while(copied < size);
 616         (*pos) += size;
 617         return size;
 618 }
 619
 620 /**
 621  * gfs2_readpages - Read a bunch of pages at once
 622  * @file: The file to read from
 623  * @mapping: Address space info
 624  * @pages: List of pages to read
 625  * @nr_pages: Number of pages to read
 626  *
 627  * Some notes:
 628  * 1. This is only for readahead, so we can simply ignore any things
 629  *    which are slightly inconvenient (such as locking conflicts between
 630  *    the page lock and the glock) and return having done no I/O. Its
 631  *    obviously not something we'd want to do on too regular a basis.
 632  *    Any I/O we ignore at this time will be done via readpage later.
 633  * 2. We don't handle stuffed files here we let readpage do the honours.
 634  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 635  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
 636  */
 637
 638 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 639                           struct list_head *pages, unsigned nr_pages)
 640 {
 641         struct inode *inode = mapping->host;
 642         struct gfs2_inode *ip = GFS2_I(inode);
 643         struct gfs2_sbd *sdp = GFS2_SB(inode);
 644         struct gfs2_holder gh;
 645         int ret;
 646
 647         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
 648         ret = gfs2_glock_nq(&gh);
 649         if (unlikely(ret))
 650                 goto out_uninit;
 651         if (!gfs2_is_stuffed(ip))
 652                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
 653         gfs2_glock_dq(&gh);
 654 out_uninit:
 655         gfs2_holder_uninit(&gh);
 656         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 657                 ret = -EIO;
 658         return ret;
 659 }
 660
 661 /**
 662  * gfs2_write_begin - Begin to write to a file
 663  * @file: The file to write to
 664  * @mapping: The mapping in which to write
 665  * @pos: The file offset at which to start writing
 666  * @len: Length of the write
 667  * @flags: Various flags
 668  * @pagep: Pointer to return the page
 669  * @fsdata: Pointer to return fs data (unused by GFS2)
 670  *
 671  * Returns: errno
 672  */
 673
 674 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
 675                             loff_t pos, unsigned len, unsigned flags,
 676                             struct page **pagep, void **fsdata)
 677 {
 678         struct gfs2_inode *ip = GFS2_I(mapping->host);
 679         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
 680         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 681         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
 682         unsigned requested = 0;
 683         int alloc_required;
 684         int error = 0;
 685         pgoff_t index = pos >> PAGE_SHIFT;
 686         unsigned from = pos & (PAGE_SIZE - 1);
 687         struct page *page;
 688
 689         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
 690         error = gfs2_glock_nq(&ip->i_gh);
 691         if (unlikely(error))
 692                 goto out_uninit;
 693         if (&ip->i_inode == sdp->sd_rindex) {
 694                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
 695                                            GL_NOCACHE, &m_ip->i_gh);
 696                 if (unlikely(error)) {
 697                         gfs2_glock_dq(&ip->i_gh);
 698                         goto out_uninit;
 699                 }
 700         }
 701
 702         alloc_required = gfs2_write_alloc_required(ip, pos, len);
 703
 704         if (alloc_required || gfs2_is_jdata(ip))
 705                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
 706
 707         if (alloc_required) {
 708                 struct gfs2_alloc_parms ap = { .aflags = 0, };
 709                 requested = data_blocks + ind_blocks;
 710                 ap.target = requested;
 711                 error = gfs2_quota_lock_check(ip, &ap);
 712                 if (error)
 713                         goto out_unlock;
 714
 715                 error = gfs2_inplace_reserve(ip, &ap);
 716                 if (error)
 717                         goto out_qunlock;
 718         }
 719
 720         rblocks = RES_DINODE + ind_blocks;
 721         if (gfs2_is_jdata(ip))
 722                 rblocks += data_blocks ? data_blocks : 1;
 723         if (ind_blocks || data_blocks)
 724                 rblocks += RES_STATFS + RES_QUOTA;
 725         if (&ip->i_inode == sdp->sd_rindex)
 726                 rblocks += 2 * RES_STATFS;
 727         if (alloc_required)
 728                 rblocks += gfs2_rg_blocks(ip, requested);
 729
 730         error = gfs2_trans_begin(sdp, rblocks,
 731                                  PAGE_SIZE/sdp->sd_sb.sb_bsize);
 732         if (error)
 733                 goto out_trans_fail;
 734
 735         error = -ENOMEM;
 736         flags |= AOP_FLAG_NOFS;
 737         page = grab_cache_page_write_begin(mapping, index, flags);
 738         *pagep = page;
 739         if (unlikely(!page))
 740                 goto out_endtrans;
 741
 742         if (gfs2_is_stuffed(ip)) {
 743                 error = 0;
 744                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 745                         error = gfs2_unstuff_dinode(ip, page);
 746                         if (error == 0)
 747                                 goto prepare_write;
 748                 } else if (!PageUptodate(page)) {
 749                         error = stuffed_readpage(ip, page);
 750                 }
 751                 goto out;
 752         }
 753
 754 prepare_write:
 755         error = __block_write_begin(page, from, len, gfs2_block_map);
 756 out:
 757         if (error == 0)
 758                 return 0;
 759
 760         unlock_page(page);
 761         put_page(page);
 762
 763         gfs2_trans_end(sdp);
 764         if (pos + len > ip->i_inode.i_size)
 765                 gfs2_trim_blocks(&ip->i_inode);
 766         goto out_trans_fail;
 767
 768 out_endtrans:
 769         gfs2_trans_end(sdp);
 770 out_trans_fail:
 771         if (alloc_required) {
 772                 gfs2_inplace_release(ip);
 773 out_qunlock:
 774                 gfs2_quota_unlock(ip);
 775         }
 776 out_unlock:
 777         if (&ip->i_inode == sdp->sd_rindex) {
 778                 gfs2_glock_dq(&m_ip->i_gh);
 779                 gfs2_holder_uninit(&m_ip->i_gh);
 780         }
 781         gfs2_glock_dq(&ip->i_gh);
 782 out_uninit:
 783         gfs2_holder_uninit(&ip->i_gh);
 784         return error;
 785 }
 786
 787 /**
 788  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 789  * @inode: the rindex inode
 790  */
 791 static void adjust_fs_space(struct inode *inode)
 792 {
 793         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 794         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 795         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
 796         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 797         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 798         struct buffer_head *m_bh, *l_bh;
 799         u64 fs_total, new_free;
 800
 801         /* Total up the file system space, according to the latest rindex. */
 802         fs_total = gfs2_ri_total(sdp);
 803         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
 804                 return;
 805
 806         spin_lock(&sdp->sd_statfs_spin);
 807         gfs2_statfs_change_in(m_sc, m_bh->b_data +
 808                               sizeof(struct gfs2_dinode));
 809         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 810                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 811         else
 812                 new_free = 0;
 813         spin_unlock(&sdp->sd_statfs_spin);
 814         fs_warn(sdp, "File system extended by %llu blocks.\n",
 815                 (unsigned long long)new_free);
 816         gfs2_statfs_change(sdp, new_free, new_free, 0);
 817
 818         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
 819                 goto out;
 820         update_statfs(sdp, m_bh, l_bh);
 821         brelse(l_bh);
 822 out:
 823         brelse(m_bh);
 824 }
 825
 826 /**
 827  * gfs2_stuffed_write_end - Write end for stuffed files
 828  * @inode: The inode
 829  * @dibh: The buffer_head containing the on-disk inode
 830  * @pos: The file position
 831  * @len: The length of the write
 832  * @copied: How much was actually copied by the VFS
 833  * @page: The page
 834  *
 835  * This copies the data from the page into the inode block after
 836  * the inode data structure itself.
 837  *
 838  * Returns: errno
 839  */
 840 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
 841                                   loff_t pos, unsigned len, unsigned copied,
 842                                   struct page *page)
 843 {
 844         struct gfs2_inode *ip = GFS2_I(inode);
 845         struct gfs2_sbd *sdp = GFS2_SB(inode);
 846         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 847         u64 to = pos + copied;
 848         void *kaddr;
 849         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
 850
 851         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
 852         kaddr = kmap_atomic(page);
 853         memcpy(buf + pos, kaddr + pos, copied);
 854         flush_dcache_page(page);
 855         kunmap_atomic(kaddr);
 856
 857         WARN_ON(!PageUptodate(page));
 858         unlock_page(page);
 859         put_page(page);
 860
 861         if (copied) {
 862                 if (inode->i_size < to)
 863                         i_size_write(inode, to);
 864                 mark_inode_dirty(inode);
 865         }
 866
 867         if (inode == sdp->sd_rindex) {
 868                 adjust_fs_space(inode);
 869                 sdp->sd_rindex_uptodate = 0;
 870         }
 871
 872         brelse(dibh);
 873         gfs2_trans_end(sdp);
 874         if (inode == sdp->sd_rindex) {
 875                 gfs2_glock_dq(&m_ip->i_gh);
 876                 gfs2_holder_uninit(&m_ip->i_gh);
 877         }
 878         gfs2_glock_dq(&ip->i_gh);
 879         gfs2_holder_uninit(&ip->i_gh);
 880         return copied;
 881 }
 882
 883 /**
 884  * gfs2_write_end
 885  * @file: The file to write to
 886  * @mapping: The address space to write to
 887  * @pos: The file position
 888  * @len: The length of the data
 889  * @copied: How much was actually copied by the VFS
 890  * @page: The page that has been written
 891  * @fsdata: The fsdata (unused in GFS2)
 892  *
 893  * The main write_end function for GFS2. We have a separate one for
 894  * stuffed files as they are slightly different, otherwise we just
 895  * put our locking around the VFS provided functions.
 896  *
 897  * Returns: errno
 898  */
 899
 900 static int gfs2_write_end(struct file *file, struct address_space *mapping,
 901                           loff_t pos, unsigned len, unsigned copied,
 902                           struct page *page, void *fsdata)
 903 {
 904         struct inode *inode = page->mapping->host;
 905         struct gfs2_inode *ip = GFS2_I(inode);
 906         struct gfs2_sbd *sdp = GFS2_SB(inode);
 907         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
 908         struct buffer_head *dibh;
 909         unsigned int from = pos & (PAGE_SIZE - 1);
 910         unsigned int to = from + len;
 911         int ret;
 912         struct gfs2_trans *tr = current->journal_info;
 913         BUG_ON(!tr);
 914
 915         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
 916
 917         ret = gfs2_meta_inode_buffer(ip, &dibh);
 918         if (unlikely(ret)) {
 919                 unlock_page(page);
 920                 put_page(page);
 921                 goto failed;
 922         }
 923
 924         if (gfs2_is_stuffed(ip))
 925                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
 926
 927         if (!gfs2_is_writeback(ip))
 928                 gfs2_page_add_databufs(ip, page, from, to);
 929
 930         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
 931         if (tr->tr_num_buf_new)
 932                 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
 933         else
 934                 gfs2_trans_add_meta(ip->i_gl, dibh);
 935
 936
 937         if (inode == sdp->sd_rindex) {
 938                 adjust_fs_space(inode);
 939                 sdp->sd_rindex_uptodate = 0;
 940         }
 941
 942         brelse(dibh);
 943 failed:
 944         gfs2_trans_end(sdp);
 945         gfs2_inplace_release(ip);
 946         if (ip->i_qadata && ip->i_qadata->qa_qd_num)
 947                 gfs2_quota_unlock(ip);
 948         if (inode == sdp->sd_rindex) {
 949                 gfs2_glock_dq(&m_ip->i_gh);
 950                 gfs2_holder_uninit(&m_ip->i_gh);
 951         }
 952         gfs2_glock_dq(&ip->i_gh);
 953         gfs2_holder_uninit(&ip->i_gh);
 954         return ret;
 955 }
 956
 957 /**
 958  * gfs2_set_page_dirty - Page dirtying function
 959  * @page: The page to dirty
 960  *
 961  * Returns: 1 if it dirtyed the page, or 0 otherwise
 962  */
 963
 964 static int gfs2_set_page_dirty(struct page *page)
 965 {
 966         SetPageChecked(page);
 967         return __set_page_dirty_buffers(page);
 968 }
 969
 970 /**
 971  * gfs2_bmap - Block map function
 972  * @mapping: Address space info
 973  * @lblock: The block to map
 974  *
 975  * Returns: The disk address for the block or 0 on hole or error
 976  */
 977
 978 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 979 {
 980         struct gfs2_inode *ip = GFS2_I(mapping->host);
 981         struct gfs2_holder i_gh;
 982         sector_t dblock = 0;
 983         int error;
 984
 985         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 986         if (error)
 987                 return 0;
 988
 989         if (!gfs2_is_stuffed(ip))
 990                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
 991
 992         gfs2_glock_dq_uninit(&i_gh);
 993
 994         return dblock;
 995 }
 996
 997 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
 998 {
 999         struct gfs2_bufdata *bd;
1000
1001         lock_buffer(bh);
1002         gfs2_log_lock(sdp);
1003         clear_buffer_dirty(bh);
1004         bd = bh->b_private;
1005         if (bd) {
1006                 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
1007                         list_del_init(&bd->bd_list);
1008                 else
1009                         gfs2_remove_from_journal(bh, REMOVE_JDATA);
1010         }
1011         bh->b_bdev = NULL;
1012         clear_buffer_mapped(bh);
1013         clear_buffer_req(bh);
1014         clear_buffer_new(bh);
1015         gfs2_log_unlock(sdp);
1016         unlock_buffer(bh);
1017 }
1018
1019 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
1020                                 unsigned int length)
1021 {
1022         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
1023         unsigned int stop = offset + length;
1024         int partial_page = (offset || length < PAGE_SIZE);
1025         struct buffer_head *bh, *head;
1026         unsigned long pos = 0;
1027
1028         BUG_ON(!PageLocked(page));
1029         if (!partial_page)
1030                 ClearPageChecked(page);
1031         if (!page_has_buffers(page))
1032                 goto out;
1033
1034         bh = head = page_buffers(page);
1035         do {
1036                 if (pos + bh->b_size > stop)
1037                         return;
1038
1039                 if (offset <= pos)
1040                         gfs2_discard(sdp, bh);
1041                 pos += bh->b_size;
1042                 bh = bh->b_this_page;
1043         } while (bh != head);
1044 out:
1045         if (!partial_page)
1046                 try_to_release_page(page, 0);
1047 }
1048
1049 /**
1050  * gfs2_ok_for_dio - check that dio is valid on this file
1051  * @ip: The inode
1052  * @offset: The offset at which we are reading or writing
1053  *
1054  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
1055  *          1 (to accept the i/o request)
1056  */
1057 static int gfs2_ok_for_dio(struct gfs2_inode *ip, loff_t offset)
1058 {
1059         /*
1060          * Should we return an error here? I can't see that O_DIRECT for
1061          * a stuffed file makes any sense. For now we'll silently fall
1062          * back to buffered I/O
1063          */
1064         if (gfs2_is_stuffed(ip))
1065                 return 0;
1066
1067         if (offset >= i_size_read(&ip->i_inode))
1068                 return 0;
1069         return 1;
1070 }
1071
1072
1073
1074 static ssize_t gfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
1075 {
1076         struct file *file = iocb->ki_filp;
1077         struct inode *inode = file->f_mapping->host;
1078         struct address_space *mapping = inode->i_mapping;
1079         struct gfs2_inode *ip = GFS2_I(inode);
1080         loff_t offset = iocb->ki_pos;
1081         struct gfs2_holder gh;
1082         int rv;
1083
1084         /*
1085          * Deferred lock, even if its a write, since we do no allocation
1086          * on this path. All we need change is atime, and this lock mode
1087          * ensures that other nodes have flushed their buffered read caches
1088          * (i.e. their page cache entries for this inode). We do not,
1089          * unfortunately have the option of only flushing a range like
1090          * the VFS does.
1091          */
1092         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1093         rv = gfs2_glock_nq(&gh);
1094         if (rv)
1095                 goto out_uninit;
1096         rv = gfs2_ok_for_dio(ip, offset);
1097         if (rv != 1)
1098                 goto out; /* dio not valid, fall back to buffered i/o */
1099
1100         /*
1101          * Now since we are holding a deferred (CW) lock at this point, you
1102          * might be wondering why this is ever needed. There is a case however
1103          * where we've granted a deferred local lock against a cached exclusive
1104          * glock. That is ok provided all granted local locks are deferred, but
1105          * it also means that it is possible to encounter pages which are
1106          * cached and possibly also mapped. So here we check for that and sort
1107          * them out ahead of the dio. The glock state machine will take care of
1108          * everything else.
1109          *
1110          * If in fact the cached glock state (gl->gl_state) is deferred (CW) in
1111          * the first place, mapping->nr_pages will always be zero.
1112          */
1113         if (mapping->nrpages) {
1114                 loff_t lstart = offset & ~(PAGE_SIZE - 1);
1115                 loff_t len = iov_iter_count(iter);
1116                 loff_t end = PAGE_ALIGN(offset + len) - 1;
1117
1118                 rv = 0;
1119                 if (len == 0)
1120                         goto out;
1121                 if (test_and_clear_bit(GIF_SW_PAGED, &ip->i_flags))
1122                         unmap_shared_mapping_range(ip->i_inode.i_mapping, offset, len);
1123                 rv = filemap_write_and_wait_range(mapping, lstart, end);
1124                 if (rv)
1125                         goto out;
1126                 if (iov_iter_rw(iter) == WRITE)
1127                         truncate_inode_pages_range(mapping, lstart, end);
1128         }
1129
1130         rv = __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev, iter,
1131                                   gfs2_get_block_direct, NULL, NULL, 0);
1132 out:
1133         gfs2_glock_dq(&gh);
1134 out_uninit:
1135         gfs2_holder_uninit(&gh);
1136         return rv;
1137 }
1138
1139 /**
1140  * gfs2_releasepage - free the metadata associated with a page
1141  * @page: the page that's being released
1142  * @gfp_mask: passed from Linux VFS, ignored by us
1143  *
1144  * Call try_to_free_buffers() if the buffers in this page can be
1145  * released.
1146  *
1147  * Returns: 0
1148  */
1149
1150 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1151 {
1152         struct address_space *mapping = page->mapping;
1153         struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
1154         struct buffer_head *bh, *head;
1155         struct gfs2_bufdata *bd;
1156
1157         if (!page_has_buffers(page))
1158                 return 0;
1159
1160         /*
1161          * From xfs_vm_releasepage: mm accommodates an old ext3 case where
1162          * clean pages might not have had the dirty bit cleared.  Thus, it can
1163          * send actual dirty pages to ->releasepage() via shrink_active_list().
1164          *
1165          * As a workaround, we skip pages that contain dirty buffers below.
1166          * Once ->releasepage isn't called on dirty pages anymore, we can warn
1167          * on dirty buffers like we used to here again.
1168          */
1169
1170         gfs2_log_lock(sdp);
1171         spin_lock(&sdp->sd_ail_lock);
1172         head = bh = page_buffers(page);
1173         do {
1174                 if (atomic_read(&bh->b_count))
1175                         goto cannot_release;
1176                 bd = bh->b_private;
1177                 if (bd && bd->bd_tr)
1178                         goto cannot_release;
1179                 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
1180                         goto cannot_release;
1181                 bh = bh->b_this_page;
1182         } while(bh != head);
1183         spin_unlock(&sdp->sd_ail_lock);
1184
1185         head = bh = page_buffers(page);
1186         do {
1187                 bd = bh->b_private;
1188                 if (bd) {
1189                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1190                         if (!list_empty(&bd->bd_list))
1191                                 list_del_init(&bd->bd_list);
1192                         bd->bd_bh = NULL;
1193                         bh->b_private = NULL;
1194                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1195                 }
1196
1197                 bh = bh->b_this_page;
1198         } while (bh != head);
1199         gfs2_log_unlock(sdp);
1200
1201         return try_to_free_buffers(page);
1202
1203 cannot_release:
1204         spin_unlock(&sdp->sd_ail_lock);
1205         gfs2_log_unlock(sdp);
1206         return 0;
1207 }
1208
1209 static const struct address_space_operations gfs2_writeback_aops = {
1210         .writepage = gfs2_writepage,
1211         .writepages = gfs2_writepages,
1212         .readpage = gfs2_readpage,
1213         .readpages = gfs2_readpages,
1214         .write_begin = gfs2_write_begin,
1215         .write_end = gfs2_write_end,
1216         .bmap = gfs2_bmap,
1217         .invalidatepage = gfs2_invalidatepage,
1218         .releasepage = gfs2_releasepage,
1219         .direct_IO = gfs2_direct_IO,
1220         .migratepage = buffer_migrate_page,
1221         .is_partially_uptodate = block_is_partially_uptodate,
1222         .error_remove_page = generic_error_remove_page,
1223 };
1224
1225 static const struct address_space_operations gfs2_ordered_aops = {
1226         .writepage = gfs2_writepage,
1227         .writepages = gfs2_writepages,
1228         .readpage = gfs2_readpage,
1229         .readpages = gfs2_readpages,
1230         .write_begin = gfs2_write_begin,
1231         .write_end = gfs2_write_end,
1232         .set_page_dirty = gfs2_set_page_dirty,
1233         .bmap = gfs2_bmap,
1234         .invalidatepage = gfs2_invalidatepage,
1235         .releasepage = gfs2_releasepage,
1236         .direct_IO = gfs2_direct_IO,
1237         .migratepage = buffer_migrate_page,
1238         .is_partially_uptodate = block_is_partially_uptodate,
1239         .error_remove_page = generic_error_remove_page,
1240 };
1241
1242 static const struct address_space_operations gfs2_jdata_aops = {
1243         .writepage = gfs2_jdata_writepage,
1244         .writepages = gfs2_jdata_writepages,
1245         .readpage = gfs2_readpage,
1246         .readpages = gfs2_readpages,
1247         .write_begin = gfs2_write_begin,
1248         .write_end = gfs2_write_end,
1249         .set_page_dirty = gfs2_set_page_dirty,
1250         .bmap = gfs2_bmap,
1251         .invalidatepage = gfs2_invalidatepage,
1252         .releasepage = gfs2_releasepage,
1253         .is_partially_uptodate = block_is_partially_uptodate,
1254         .error_remove_page = generic_error_remove_page,
1255 };
1256
1257 void gfs2_set_aops(struct inode *inode)
1258 {
1259         struct gfs2_inode *ip = GFS2_I(inode);
1260
1261         if (gfs2_is_writeback(ip))
1262                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1263         else if (gfs2_is_ordered(ip))
1264                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1265         else if (gfs2_is_jdata(ip))
1266                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1267         else
1268                 BUG();
1269 }
1270